Replit CEO Amjad Masad
Amjad Masad argues that even if frontier research stopped today, existing AI capabilities would take another 10 years to diffuse throughout the economy. He points to solution partners as a way to accelerate that process, emphasizing the implementation work still ahead. The timeline is his assessment of adoption.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux’s day-nine update focuses on optimizing disk-state storage in the ChatGPT and Codex desktop apps. He expects the optimization to reduce the relevant disk usage by 60%. He especially wants feedback from power users about how much space it frees in practice. This is an expected improvement awaiting user feedback.
Box CEO Aaron Levie
Aaron Levie shares Box’s internal AI Complex Work evaluation, reporting a 14-percentage-point improvement for Gemini 4 Argon over 3.1 Pro in mathematics, logic, domain knowledge and following tasks through to completion. Sector gains are 18 points in finance, 25 in healthcare, 37 in technology, 16 in energy and 33 in consumer businesses, all from Box’s own evaluation. The finance case requires annual bank-loan and overdraft interest in a funding-gap model; in healthcare, both models identify dangerous potassium levels and drug interactions, but only Argon removes the affected drugs from the final regimen. The technology case tests monthly, regional retention and per-user metrics; the energy case detects meter records ending in November; the consumer case uses inventory to bridge an early capacity gap and assesses expansion and launch timing. He says Argon will become available in Box AI Studio. In another long post, he argues that agents will greatly expand the attack surface while creating opportunities for cybersecurity companies, describing the expansion as “100fold.” He envisions agent swarms reviewing code, assessing combinations of risks, responding to incidents and clearing security backlogs across model labs, networks, firewalls, identity, endpoints, software development and data. That attack-surface multiplier is his judgment, not a measurement from the evaluation.
Google vice president Josh Woodward
Josh Woodward shows two uses of Gemini Argon. One designs a custom 96mm furniture leg and explores the number of stacked spheres in the design. The other uses a 24-word prompt for a synthesizer combining galloping horses, water and hand movements into a musical experience. The demonstrations concern dimension-specific design and interactive creation from a short input; the posts do not supply systematic evaluation results.
Y Combinator CEO Garry Tan
Garry Tan argues that connecting an agent harness to a real browser and real credentials unlocks new capabilities. He recommends Aside AI, focusing on agents working in the environments people actually use. The post does not elaborate on implementation or permission configuration.
Vercel CEO Guillermo Rauch
Guillermo Rauch congratulates Gemini’s progress, pointing to Google’s transformer heritage, data, engineers and compute. He sees coding agents as a key application. The arrival of Muse Wonder and Gemini Argon in quick succession also reinforces his interest in model choice. Vercel will remain model-agnostic through Gateway and fx. His focus is giving developers access to advances from different labs.
AI Engineer and Latent Space affiliate Swyx
Swyx shares the “How to Kill Your SaaS” challenge, in which participants attempt to clone actual SaaS costing $43,000 a year over a weekend. The global hackathon received 72 submissions and used approximately 100 billion tokens. Evaluation combines AI and human review, with $10,000 in prizes. His focus is the ability to recreate real software and the evaluation process. This item does not name the winners, so it does not establish a ranking.
Anthropic Engineering: Redesigning source control at Anthropic
Anthropic has redesigned source control for an enormous monorepo and agent-generated code volumes, migrating to its internal Scion system alongside the Antique code-review tool. It reports daily merges rising from about 2,500 in May to over 30,000 in September, with CI performing 25 million checkouts a day. Scion moves history queries and operations such as merging to servers. Clients request immutable trees and blobs as needed, with node and regional caches each achieving hit rates above 99%. According to internal results, CPU and memory infrastructure requirements are 200 times smaller, checkout takes seconds instead of minutes, and CI gains 25% CPU capacity. Seven people built the systems and completed migration in five months. A Claude Mythos Preview rewrite of an ingestion pipeline used 20 times less memory and ran four times faster. The article stresses: “Engineers are no less valuable, but their work is changing.” Engineers still own system constraints, tradeoffs and intent review, stopping agents from breaking constraints or overengineering. Scion is not a general-purpose git replacement: migration depends on accompanying review, search and CI tools, and Anthropic’s internal token spending does not represent a cost ordinary businesses could reasonably bear.
Training Data — David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
**Key takeaway: Ineffable Intelligence aims to explore a superlearner that discovers knowledge from experience, rather than extending human text into a larger library of answers.** David Silver and Lasse Espeholt founded Ineffable Intelligence, building on their reinforcement-learning experience at DeepMind; Espeholt worked on IMPALA and SEED RL. They define the goal in terms of depth and breadth: discovering things beyond existing human knowledge and learning from simulated, digital and physical experience. First, simple objectives can generate complex learning. Silver uses Go to illustrate the point: a lone player on a 19-by-19 board has limited possibilities, but an opponent continually creates new strategic problems. They want to build such “crucibles,” where physical rules and multi-agent interactions generate challenges instead of a hand-authored trail of puzzles. Silver summarizes: “Everything we learned, we discovered for ourselves.” Second, simulation offers speed and reproducibility. Large volumes of parallel experience can reduce the cost of real-world trial and error; simulations running 10 to 100 times faster would create more room for learning. They focus on transferable learning algorithms, without equating success in simulation with reliable deployment. They do not plan an intermediate LLM product stage. Third, safety cannot rely on verbal explanations alone. They propose shadow modes that observe actions without executing them, and supervision by other intelligent systems; these remain research directions requiring evidence. Safety validation must constrain entry into the physical world, rather than a plausible-looking chain of reasoning. Fourth, systems design serves the scale of experience. They plan to use Vera Rubin GPU clusters on Google Cloud, keeping simulation, inference and training on GPUs where possible instead of relying on enormous CPU fleets. Continuously changing reinforcement-learning data also creates scaling dynamics different from fixed-corpus training. The roughly 20-person team emphasizes running many experiments per researcher. Silver’s expectation of evidence within a year and genuine superintelligence in about three years is a forecast. The central bet is whether learning can continually create new experience; experiments still have to establish the results.